Forensic Chemistry 3 (2017) 113

Contents lists available at ScienceDirect

Forensic Chemistry
journal homepage: www.elsevier.com/locate/forc

Full length articles

Local Ion Signatures (LIS) for the examination of comprehensive

two-dimensional gas chromatography applied to fire debris analysis
Martin Lopatka a,f, Andjoe A. Sampat b, Steffan Jonkers b, Lawrence A. Adutwum c, Hans G.J. Mol d,
Guido van der Weg d, James J. Harynuk c, Peter J. Schoenmakers b, Arian van Asten b,e,f,,
Marjan J. Sjerps a,f,e, Gabriel Viv-Truyols b
a
Korteweg de Vries Institute for Mathematics, University of Amsterdam, Postbus 94248, 1090 GE Amsterdam, The Netherlands
b
Van t Hoff Institute for Molecular Sciences, University of Amsterdam, Postbus 94157, 1090 GD Amsterdam, The Netherlands
c
Department of Chemistry, University of Alberta, Gunning/Lemieux Chemistry Centre, 11227 Saskatchewan Drive, T6G 2G2 Edmonton, Alberta, Canada
d
RIKILT Wageningen UR, Postbus 230, 6700 AE Wageningen, The Netherlands
e
Co van Ledden Hulsebosch Center (CLHC), Amsterdam Center for Forensic Science and Medicine, Postbus 94157, 1090 GD Amsterdam, The Netherlands
f
Netherlands Forensic Institute, Postbus 24044, 2490 AA Den Haag, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: Forensic examination of fire debris evidence is a notoriously difficult analytical task due to the complex-
Available online 29 October 2016 ity and variability of sample composition. The use of comprehensive two-dimensional gas chromatogra-
phy with mass spectrometry detection (GC
GCMS) allows the coupling of orthogonal retention
Keywords: mechanisms and therefore a high peak capacity.
Local Ion Signature (LIS) We demonstrate recent innovations in combining chemometric techniques for data reduction and fea-
Fire Debris Analysis (FDA) ture selection, with evaluation of the evidence for forensic questions pertaining to the detection and sub-
Comprehensive two-dimensional gas
sequent classification of ignitable liquid residue (ILR) in fire debris samples. Chromatograms are divided
chromatography (GC
GCMS)
Likelihood ratio (LR)
into non-overlapping spatially delimited regions; for each of these regions a Local Ion Signature (LIS) is
Chemometrics computed by summing the intensities, per nominal mass/charge over all points contained within each
region. This yields a reduced feature space representing the original data as a set of consolidated ion
traces. Subsequent feature selection is performed by evaluating the individual efficacy of each feature
using a univariate score-based likelihood ratio (LR) approach for discriminating between pairs of same
or different type samples. The retained features are used to model each ILR class using linear discriminant
analysis (LDA).
Results are demonstrated for 155 arson samples containing a diversity of substrate compounds and
several known ignitable liquids. ILR detection is performed at 84% accuracy with fewer than 1% false pos-
itives followed by subsequent classification. Likelihood ratio distributions are presented referring to both
detection and classification tasks.
2016 Elsevier B.V. All rights reserved.

1. Introduction
1.1. Forensic relevance
Records kept by the Dutch Statistics Bureau (CBS) [1] indicate
that 19.2% of the indoor fires responded to in 2013 were suspected
Corresponding author at: Van t Hoff Institute for Molecular Sciences, University
of Amsterdam, Postbus 94157, 1090 GD Amsterdam, The Netherlands.
E-mail addresses: m.lopatka@uva.nl (M. Lopatka), a.a.s.sampat@uva.nl
(A.A. Sampat), steffan.jonkers@student.uva.nl (S. Jonkers), adutwum@ualberta.ca
(L.A. Adutwum), hans.mol@wur.nl (H.G.J. Mol), guido.vanderweg@wur.nl
(G. van der Weg), james.harynuk@ualberta.ca (J.J. Harynuk), p.j.schoenmakers@
uva.nl (P.J. Schoenmakers), a.c.vanasten@uva.nl (A. van Asten), m.j.sjerps@uva.nl
(M.J. Sjerps), g.vivotruyols@uva.nl (G. Viv-Truyols).
(by fire department first responders) to be arson cases. Due to the
immense societal cost inflicted by accidental and deliberate fires,
substantial resources are invested in the forensic analysis of fire
debris; a task requiring both forensic expertise and the capacity
to perform sophisticated chemical analyses. Forensic fire investiga-
tion may involve the detection of ignitable liquid residues (ILR) in
fire debris from the scene. Depending on the circumstances, such
residues may constitute important evidence that arson was com-
mitted and can provide valuable tactical information.
Despite the maturity of Fire Debris Analysis (FDA) as a domain
of forensic practice, detection and classification of ILR remains a
very challenging task. This detection and classification process is
complex in nature and requires extensive experience from the
forensic chemist to carry out interpretation [2]. Several factors

http://dx.doi.org/10.1016/j.forc.2016.10.003
2468-1709/ 2016 Elsevier B.V. All rights reserved.
2 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
complicate the analysis of fire debris, such as the complex compo-
sition of most ignitable liquids [3], the diversity of ignitable liquid
products easily accessible to arsonists [4,5], the partial consump-
tion and evaporation of ignitable liquids during the fire [6,7], the
formation of interfering pyrolysis products from the substrate
[8,9], the invasive processes employed by fire fighters to stop and
extinguish the fire [10] and the physical, chemical and biochemical
processes that can occur after the fire [11,12].
It has been shown that these complicating factors can lead
forensic experts to erroneous conclusions for challenging fire deb-
ris samples [13]. The subjective nature of the data interpretation
has serious implications for accreditation and harmonization
efforts as well as for the experimental design of studies use to eval-
uate the efficacy of automated methods. In the study presented
herein, the authors choose to equate the pre-burn addition of igni-
table liquids (IL) to a sample with the ground truth presence of
ignitable liquid residue (ILR) in the fire debris. The goal of this
manuscript is to demonstrate preliminary steps towards a method-
ology for objective and automated detection and classification of
ILR in fire debris samples using comprehensive two-dimensional
gas chromatography. Continued development of such methods
could assist forensic experts with data interpretation and improve
the objectivity of fire debris analysis.
1.2. Background
Previous work has demonstrated that the Total Ion Spectrum
(TIS) [14,15] calculated over an entire chromatographic separation
may yield a very valuable feature space in which the detection and
classification of fire debris can be successfully performed [1618].
The TIS is computed by summing the intensity of each nominal
mass-to-charge value over all measurements along the retention
time axis of a chromatographic separation. This approach provides
many desirable characteristics for the comparison of chromato-
graphic signals. One advantage is the robustness of this method
to retention time shifts, frequently observed in chromatographic
data [15]. Another advantage is the inherent reduction in dimen-
sionality that comes from aggregating the mass intensities.
Within the domain of FDA, many state of the art techniques
demonstrate that the TIS feature space can yield impressive results
when combined with chemometric techniques. Classification of ILR
in debris by soft Bayesian classification [19] and Principal Compo-
nent Analysis (PCA) with Linear and Quadratic Discriminant Anal-
ysis (LDA/QDA) [16] demonstrate classification into the ASTM [20]
defined IL classes. The need for dimension reduction is addressed
in these publications, specifically in [16] where a step-wise classi-
fication method is adopted as insufficient data is available for a
multi-class classification. Currently published techniques achieve
correct classification rates of 7090% with false positive error rates
of less than 10%. Although these results are impressive, the mini-
mization of false positive rates is of paramount importance in a
forensic context.
While much of the previous work in automated fire debris clas-
sification has been performed using one-dimensional chromato-
graphic data [1619,21,22] there are several examples of
analytical strategies showing promise in the progression to
comprehensive two-dimensional chromatography as a viable Fire
Debris Analysis technique [2326,21]. The use of GC
GC provides
a vast separation space in which to detect features of forensic rel-
evance for the identification and subsequent classification of igni-
table liquid residue in fire debris samples. This task often entails
computationally complex procedures, encumbered by many well
documented complications that may arise from retention time
shifting in both chromatographic dimensions [27,28], substantial
noise contribution [29], and potential obfuscation of trace com-
pound peaks even in a very large separation space [30,31].
This manuscript presents preliminary research towards the
adaptation of a TIS method to local regions in the vast separation
space afforded by a GC
GC separation. The same principles that
allow excellent performance to be accomplished using the TIS fea-
ture space in a GCMS separation with substantial peak coelution
are expected to yield an informative reduce feature space in this
context.
1.3. Interpretation of evidence
The likelihood ratio framework for quantitative assessment of
evidential value of physicochemical traces is increasingly sup-
ported in forensic research [3234], as a reporting method for
the courtroom [35,36], and as an investigative tool for forensic
intelligence [37,38]. The observation of the chemical evidence
under each of two (or more) competing hypotheses allows compar-
ison of new forensic samples in the context of information gained
from reference samples from a relevant forensic population. Two
strategies for the use of likelihood ratio models that are most fre-
quently encountered in forensic contexts are score-based and
feature-based approaches [34,39,7]. Each characterizes the evi-
dence in a distinctly different manner, making different modeling
assumptions in addressing source identity problems. Score-based
methods report on the rarity of observing a particular score
between a pair of samples, thus reformulating the evidence in
terms of the rarities of observing similarity metric comparing sam-
ples rather than the original features [34]. Feature-based LRs report
on the rarity of observing particular features under competing
hypotheses pertaining to the source of the sample [35]. Recent
research has drawn attention to the fundamental differences in
the statistical models that can be defined within the LR framework
[39,40,41] and the practical differences observed with real chemi-
cal measurements [34,42].
The use of an LR framework in the assessment of samples
requires the formulation of competing hypotheses specific to the
level of forensic interest being examined. For detection of ILR (for
an arbitrary sample, x) the hypotheses may be formulated as
follows:
Hp : sample x contains ignitable liquid residue
Hd : sample x does not contain ignitable liquid residue
Conventionally the subscripts p and d are adopted to indicate
the hypotheses of the prosecution and defence respectively. For
the classification of ILR, the hypotheses are reformulated to pertain
to the classification of a sample (x) int an arbitrary class of ignitable
liquids (y) with at least 2 mutually exclusive options here denoted i
and j.
Hj : sample x contains ILR belonging to the class yi
Hj : sample x contains ILR belonging to the class yj
The use of an LR framework for assessing fire debris samples has
the distinct advantage that it can not only be used to perform hard
classifications (if required) but also to explicitly indicate the
strength of the evidence on a quantitative scale. This is a very impor-
tant advantage for forensic casework, where experts should refrain
from making absolute claims pertaining to class membership.
The forensic context explored herein is the detection and subse-
quent classification of ignitable liquid residues into three classes of
interest: White Spirits (WS), Gasolines (GAS), and Lamp Oils (LO).
GC
GCMS is used to improve the performance (compared to
similar approaches operating on 1-D data) by providing chromato-
graphically separated mass spectra as input and performing fea-
ture selection sensitive to the elution times of compounds of
interest. This considerably extends the Total Ion Spectra (TIS)
 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
approaches explored in previously published research for one-
dimensional chromatography [15] which is used in feature reduc-
tion of our analytical measurements.
Despite a great deal of data reduction achieved by summation
across the chromatographic space (as in the TIS approach), the
dimensionality relative to the number of available samples compli-
cates statistical modeling of the original features. Research on chem-
ical profiling for source identity has demonstrated a relationship
between the evidential value computed using feature-based LRs
and the evidential value computed using source-independent
score-based LRs [34]. The score-based approach is implicitly differ-
ent than a feature-based representation of the LR wherein the rarity
of observing the features themselves is calculated under the com-
peting hypotheses [39,43]. Nonetheless, score-based LRs may act
as a proxy for the same-source versus different-source relationship
by indirectly modeling the relationship between samples via a pair-
wise similarity metric and can indicate similar forensic assertions
[34]. It is due to this relationship that the score-based method is suit-
able as a feature selection method. By modeling the LR attainable
using only a single feature at a time, a coarse estimate of that fea-
tures evidential value may be evaluated and a feature ranking
may be used for further reduction in the dimensionality of the data.
2. Materials and methods
2.1. Laboratory fire experiments
In order to produce a realistic data set in which the presence or
absence of ILR is known as well as the type of ILR that was used as
an accelerant, small-scale fire debris samples consisting of various
substrates were created in a controlled manner. Experiments were
conducted in a fire- and explosion-proof laboratory, built at the
Netherlands Forensic Institute (NFI) for hazardous experiments.
An experimental setup was designed that included a total of 155
measurements (provided as Supplementary Material). A total of 19
different substrates, consisting of furniture and building materials
commonly found in homes, were used to conduct the experiments.
Out of the 155 measurements, 110 consisted of fire debris and 45
were analyses of neat ignitable liquids. More information regarding
the brand and corresponding ASTM categories of the specific ignita-
ble liquids used can be found in the Supplementary Material.
For 20 experiments only a single type of substrate was burned,
45 experiments included a combination of substrate materials
without the addition of ignitable liquid, and 45 experiments were
performed with exactly the same substrate composition and the
addition of ignitable liquid. For each burn condition, a total of
300 grams of material was used. In each of the experiments
(excluding single substrate burns) 60% of the total mass of material
represented the primary substrate; the remaining 40% was com-
prised of 4 types of secondary substrates. Three types of ignitable
liquids were used. For each type of ignitable liquid 15 different
bottles, representing different brands, were purchased form
different retail locations throughout the Netherlands.
Burn experiments were performed on single-use, uncoated,
stainless-steel plates (345
240 mm). For experiments involving
ignitable liquid (IL), 15 mL of IL was applied in three parallel lines
using a glass Pasteur pipette. After allowing the IL to absorb for five
minutes, the samples were ignited using a gas torch. The fires were
left to burn freely for 8 min and extinguished with water using a
spray bottle. After allowing the debris to cool (approximately
30 min), the entire sample was manually collected in glass jars.
2.2. Analytical methodology
Fire debris samples were first heated for 4 h at 70 C before
sampling 0.5 mL of headspace. This headspace sample was
3
analysed according to a GC-FID screening procedure to attain a
preliminary survey of volatiles present in the samples headspace.
The maximum signal intensity observed in the screening analysis
was used to determine an appropriate headspace volume for the
extracts used in subsequent GC
GCMS analysis. These were pre-
pared by drawing headspace through a sorbent tube containing
activated charcoal using a 100 mL syringe.
GC
GCMS analyses were carried out on an Agilent Technolo-
gies 6890 N GC System with a LECO dual-stage, quad-jet thermal
modulator coupled to a Pegasus III ToF-MS (LECO, St. Joseph, MI,
USA). An Agilent DB-1 (100% dimethyl polysiloxane) first-
dimension column (30 m
250 lm i.d.; 0:5 lm film thickness)
was coupled to an Agilent DB-17 ((50%-Phenyl)-methyl polysilox-
ane) second-dimension column (1 m
100 lm i.d.; 0.2 lm film
thickness). A modulation time of 4 s was used and detection was
performed covering a 35500 m/z range.
Full details of the analytical methodology, separation program,
and suppliers of all chemical reagents can be found in the Supple-
mentary Material.
2.3. Feature reduction
Initial data processing steps are required to reduce the
dimensionality of a single measurement signal while maintaining
information that may be valuable for the forensic investigation.
The reduced form representation (RFR) entails a reduction in the
number of features (per measurement) from 800
1185
466
variables to a consolidated data space of t
466 variables where
t Local Ion Signatures (LIS) are computed from rectangular tiles
dividing the entire chromatographic space into t equally sized tiles
as depicted in Fig. 1. This approach, inspired by earlier work using
the TIS [1518], strives to include some valuable retention time
information while still benefiting from the robust and informative
reduced representation of the sample in this manner.
A LIS may be calculated for an arbitrary region in the chro-
matographic space by summing the intensity values per a partic-
ular nominal m/z over all of the points within the given region
resulting in (for the data used here) a vector of 1
466 values
for each of t regions specified. The local vectors are then concate-
nated in an order corresponding to the retention time yielding a
single vector of 1
t
466. The number of tiles t may easily
be extended to an arbitrary number of regions, keeping in mind
that smaller tile regions increase the risk of misalignment
between measurements between chromatographic runs. To
demonstrate the utility of this strategy we divide the chromato-
graphic space into only 12 large tiles (based on retention time)
meaning that the dataset resulting from all burn experiments dis-
cussed in this manuscript can be represented as a 155
5592
matrix. As seen in Fig. 1, we may capture compounds eluting at
drastically different relative retention times per dimension with
a simple grid of 4
3 tiles. It is important to note that the selec-
tion of a tiling strategy is of paramount importance and should be
considered in the context of the specific samples and include
domain knowledge of experts.
Due to the additive effect of summing over such a large number
of data points, a substantial contribution is observed in the signal
due to the accumulation of noise in the LIS, meaning that values
of 0 are not possible in the signal even when no analytes are pre-
sent. To remove some of the effects of the noise, each LIS vector
is normalized by subtracting the minimum intensity observed
per ion from the LIS as per Eq. 2, and subsequently divided
element-wise with the adjusted TIS value as per Eq. (1). Thus for
the tile t and the ion i:
0
00 list;i
list;i 0 1
tisi
4 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
Fig. 1. Local Ion Signatures (LIS) of sample #93 calculated per tile (see Supplementary Material for burn details).
where
0
 source comparisons. When the detection of ignitable liquid residue
list;i list;i  mini list;i
and tisi is the TIS calculated for all ions i over every point in the
chromatogram,
0
tisi tisi  mini tisi :
00
Each normalized LIS (list )
corresponds to a single vector of 466
quantities calculated per tile for each of t tiles that a chromatogram
is divided into. The RFR for each chromatogram i results from con-
00 00 00
catenating the list values such that: RFRi lis1 ; . . . ; lis12 . A practi-
cal limitation of the RFR is the very large discrepancy between
low abundances of trace compounds, and the abundances of very
common ions that may originate as fragmentation products of lar-
ger molecules. To reduce the effect of ions that are universally
common in all tiles, each LIS vector is subsequently divided by
the normalized TIS, see Eq. (1). The TIS is calculated as per the pro-
cedure described in [15] but summing along both retention time
dimensions. For GC
GC this equates to summing the modulated
signal. The TIS is normalized in exactly the same manner, described
in Eq. (3). The RFR of a chromatogram may thus be described as the
ratio of the local ion quantities relative to the total quantities of
those ions in the whole chromatogram. This method of normaliza-
tion has the desirable characteristic that a new sample may be
independently normalized for comparison against a reference col-
lection without depending on parameters determined from the ref-
erence collection.
2.4. Feature selection
Further dimension reduction is required to perform classification
and evidential interpretation tasks on this data set. A score-based LR
approach can be applied as a powerful tool for the selection of fea-
tures of interest. A subset of the total data reserved for parameter
tuning partition is used to independently evaluate the efficacy of
each feature in distinguishing between same-source and different-
2 is being investigated, then same-source pairwise comparisons refer
to all comparisons made between samples containing any ignitable
liquid residue of any type as well as all comparisons made between
samples not containing ignitable liquid residue. Different-source
3 comparisons refer to all comparison made between ILR present
and ILR absent samples. We may therefore formulate a univariate
score-based LR model for the general assessment of same-label ver-
sus different-label pairwise comparisons between two arbitrary
samples i and j where i j. Based on those classifications such a sys-
tem would be described by Eq. (4). Here, the evidence is simply the
absolute difference between samples, i.e. jRFRi;f  RFRj;f j mea-
sured per feature f where, f 1; . . . ; 5592. The density of the abso-
lute differences observed under each of the two competing
hypotheses (for our data) may be described by a univariate normal
probability density function g:.
gjRFRi;f  RFRj;f jjHp
LRi;j;f 4
gjRFRi;f  RFRj;f jjHd
Hp : samples RFRi and RFRj are the same label
Hd : samples RFRi and RFRj are not the same label
where label may mean ILR present/absent or a particular class of
ILR, depending on the forensic level of interest. For each variable
(f) in the RFR of the chromatogram we perform the procedure
described above. Training samples are compared based on their
absolute difference; this process is performed exhaustively for all
pairwise comparisons possible. For the forensic classification task
the sense of intra-class similarity corresponds to pairwise compar-
isons performed between samples labels in terms of their ILR con-
tent. Adopting a labeling scheme based on the specific types of
ignitable liquids: WS, GAS, and LO, any comparisons between a
GAS sample and another GAS sample contributes to the
distribution of same-label comparisons. Likewise, comparisons
 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
between different class ignitable liquids (including comparisons
with ILR absent samples) count towards the different-label compar-
isons when generating the score-based LR model.
This entire procedure is repeated once per feature in the LIS rep-
resentation, leading to pairs of univariate distributions describing
the observed absolute differences between samples of the same
and of different type, for only one feature at a time. This means that
5592 score-based LR systems have been constructed for the detec-
tion of ILR and another 5592 for the forensic task of ILR classifica-
tion. Having also set aside a data partition for the tuning of
parameters, the performance of these LR systems in distinguishing
between similar label and dissimilar label measurements with a
known ground truth is evaluated. Using the known ground truth
and the calculated LRs it is possible to evaluate the efficacy of each
variable (f) in distinguishing between types. The metric by which
each feature is examined is the log likelihood ratio cost or CLLR
[41], calculated by Eq. (5):
  ! available to reliably calculate the covariance parameter needed in
1 1 X 1 X

N SL N DL
1 the subsequent processing steps. Therefore, the choice of 25 fea-
C llr log 2 1 log 2 1 LRDLi
2 NSL i1 LRSLi NDL i1
where NSL is the number of comparisons made between samples
with the same label and N DL is the number of different-label com-
parisons performed. LRSL and LRDL are the corresponding likelihood
ratios for same-label and different-label comparisons respectively.
The CLLR metric imposes a penalty based on the magnitude of LRs
inconsistent with the known ground truth [41]. Deviation from a
CLLR of 0 is indicative of poor performance of a system for reporting
evidence. The CLLR accounts for not only the incidence of mislead-
ing evidence but also the magnitude of the erroneous LRs reported.
In this case, each system of LRs being evaluated comprises of a sin-
gle feature in the RFR feature space.
Thus the evaluation of potential utility for each feature is based
on the CLLR achieved. Fig. 2 shows the CLLRs for an arbitrary Monte
Carlo iteration. The 5592 features in the RFR are listed along the x-
axis. For the stratified CV conditions used here 10 data partitions
are defined, therefore the parameter tuning partition consists of
15 samples and their corresponding 105 pairwise comparisons
are used in computing a single CLLR per feature. The features are
now ranked based on their corresponding CLLR value and the top
ranked features (i.e. those with the smallest CLLR values) are
retained to use in subsequent modeling steps. Any subset of the
features selected may be directly related to both a spatial position
in the original two-dimensional chromatogram (specific to a tile)
and a specific nominal m/z value. This is a valuable characteristic
Fig. 2. CLLR value calculated per feature in reduced LIS representation of sample #93.
5
when interpreting the evidential value of any assertions made on
the basis of these features and maintains a clear and intuitive rela-
tionship with the chemical information contained in the original
data (See discussion in Section 4.1). The application of a CLLR mea-
surement on each individual feature does not address the multi-
collinearity problem frequently encountered when performing
statistical analysis on chemical trace data [44]. Nonetheless, the
generalization into a series of univariate score-based LR models
allows a pragmatic basis for further feature selection.
2.5. Ignitable liquid detection and classification
A classification subspace is generated by discarding features not
ranked in the top 25 positions by the procedure described in Sec-
tion 2.4. The number of features used must strike a fine balance
in order to retain the maximum amount of information possible
from the original data while ensuring that sufficient samples are
5
tures in our case is dictated by the number of samples available.
The top 25 features of the training partition data are then projected
onto a subspace defined by performing LDA [44]. If the forensic
level of interest pertains to ignitable liquid detection, the compet-
ing hypotheses may be represented as two classes. Ground truth
labeling of each measurement is performed according to the pres-
ence or absence of ignitable liquid residue.
yi 1 if no ignitable liquid was used in the experimental burn.
yi 2 if any type of ignitable liquid was used as an accelerant in
the experimental burn.
LDA approaches a two-class classification problem by approxi-
mating a conditional probability density function for each class
under the assumption that the two classes have identical covari-
ance parameters and distinct means and that both are normally
distributed [44].
The classification of a new measurement (RFRnew ) is a function
of its position in the projected space relative to some decision
boundary. Under the presupposition of equal variance the class
membership of a new point is simply defined as the minimum dis-
tance to a class mean in the projected space. In the application
herein we utilize the LDA projection as a coarse method in order
to compute an ideal separation hyperplane using the top 25 fea-
tures previously selected by the method described in Section 2.5.
By definition, the LDA projection will be a linear combination of
6 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
Fig. 3. 1-D projection (LDA axis) density map for ILR present (red) and ILR absent (blue) classes and new points to be classified.
the features selected by their CLLR ranking and allows for a further
reduction in dimensionality while minimizing the information loss
incurred. Furthermore, the multicollinearity among these features
is accounted for by their relative contribution to the projected
space. The projected training samples are then modeled by a
Gaussian probability density functions describing the different
classes from the projected data without the added constraint of
homoscedasticity [45].
An example depicting the projection of new data points to be
classified, in terms of the two-class ignitable liquid detection task,
can be seen in Fig. 3. Since the reduced dimensional hyperplane is
defined in terms of a linear combination of the original features,
the ratio of the densities between the two classes may also be
expressed in terms of a feature-based LR, wherein the evidential
term is now the representation of a new measurement in the
one-dimensional feature space defined by the LDA projection for
any given data point. The LR for a given classification can then be
expressed by Eq. (6).
 
g projectionRFRi jly1 ; ry1
LR   6
g projectionRFRi jly2 ; ry2
The competing hypotheses are now represented by the param-
eters defining a Gaussian distribution representing that class of
data points on the LDA projected line, where g: is the Gaussian
probability density function.
The extension into multi-class LDA is described in detail [44]
and is also described alongside a computational implementation
that formed the basis of the one used in the calculation of these
results [46]. Essentially, the extension can be seen as classifying
a number of samples into three distinct categories based on select-
ing a two-dimensional projection of the original data maximizing
the variance between classes while minimizing the variance within
classes [44]. The LR formulation, in analogy with the two class for-
mulation described in Eq. (6) may be reformulated as the likeli-
hood of a given point belonging to a particular class versus the
likelihood of it belonging to another specific class, as seen in Eq.
(7) for class 1 versus class 2.
 
g projectionRFRi jlyc1 ; Ryc1
LR   7
g projectionRFRi jlyc2 ; Ryc2
In the case of a three class formulation the hypotheses are for-
mulated as follows. This formulation explicitly disregards the pos-
sibility of additional classes and false positives, however, may be
extended to any number of classes, so long as they are explicitly
Table 1
Hypotheses for the classification of ignitable liquid residue.
WS Hp : sample x contains ILR of type WS
Hd1 : sample x contains ILR of type GAS
Hd2 : sample x contains ILR of type LO
GAS Hp : sample x contains ILR of type GAS
Hd1 : sample x contains ILR of type WS
Hd2 : sample x contains ILR of type LO
LO Hp : sample x contains ILR of type LO
Hd1 : sample x contains ILR of type GAS
Hd2 : sample x contains ILR of type WS
defined and mutually exclusive. For any new classification in a 3
class problem, 2 LRs are produced for the hypotheses specified in
Table 1. If one were to model the denominator as a conjunction
of Hd1 and Hd2 an implicit prior probability is unavoidably intro-
duced pertaining to the relative rarity of each IL class in the rele-
vant population. Results reported here focus on the likelihoods
when evaluating a one-versus-one formulation of the competing
hypotheses.
In the LDA subspace, the likelihood of observing the evidence
(projected features values) under each of the Hd cases must be
computed. An example, representing new samples projected into
a two-dimensional hyperplane for the three-class ignitable liquid
residue classification problem is depicted in Fig. 4.
3. Results
In order to guard against overfitting, all results reported are
based on aggregate findings in a series of repeated Monte Carlo
iterations using a stratified cross-validation (CV) design with 10
folds. The data is partitioned such that every Monte Carlo iteration
contains the full 10 step cross-validation run where: 1=10 of the
data (proportionally sampled per experiment type and ignitable
liquid type) is assigned to be a testing partition, 1=10 is assigned
to be a parameter tuning partition and the remaining 8=10 are used
as a training partition. For one Monte Carlo iteration to complete,
each of the 10 partitions is used exactly once as the testing parti-
tion. An algorithmic representation of the Monte Carlo workflow
is presented in the Supplementary Material.
Example chromatograms from each of the three IL classes
focussed on within the context of this research are presented in
Fig. 5 demonstrating the complexity of the samples and abundance
of potential chemical information that acts as a starting point for
both manual examination by forensic chemists and automated,
statistically driven methods. By using GC
GC to drastically
 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
Fig. 4. 2-D projection (LDA axes) density maps for the 3 IL classes and new points to be classified.
increase the number of resolvable peaks in the resulting chro-
matograms an abundance of data is created to facilitate the selec-
tion of informative features for the detection and classification
tasks explored.
3.1. Detection of ignitable liquid residue
In the first step, the detection of ILR is performed. Samples from
the experimental measurements that contain only intact ILs are
removed. This is done for two reasons; firstly, post burn ILs differ
substantially from intact ILs. The inclusion of several neat IL sam-
ples in the training data partition may adversely affect the model
by allowing features to be selected that will not be present in fire
debris samples. Secondly, the focus of this study is to demonstrate
results specific to fire debris samples. Due to the use of a Monte
Carlo procedure the results presented here are aggregated over
11000 comparisons guaranteeing (by the stratified CV structure)
that each of the 110 samples has been classified as part of the test-
ing block 100 times. Hard classification is performed by threshold-
ing the resulting LR value pertaining to the hypotheses stated in
Section 2.5. A natural threshold is set at 1 where an LR > 1 provides
support in favor of Hp whereas an LR < 1 supports Hd and an LR of
exactly 1 means the evidence is equally likely under the competing
hypotheses and provides no evidential value. The confusion matrix
in Table 2 summarizes this performance. Note that the abbrevia-
tions for True Negatives (TN), False Positives (FP), False Negatives
(FN), and True Negatives (TN) along with their corresponding true
and predicted labels are introduced here.
Accurate classification performance is achieved for an aggregate
of 84% of sample comparisons (9269 out of 11000 total compar-
isons). In addition the low FP rate is a desirable trait for a forensic
application. Only samples that are selected as containing ILs are
then passed on to the classification step, including both FP and
TP samples. This method of reporting results is in line with the
development of an automated method for fire debris pre-
screening, wherein samples that are deemed to be ILR negative
are unlikely to be examined further. The number of samples that
pass from the detection step to the classification step is variable
depending on ILR detection performance for that Monte Carlo iter-
ation. The same procedure of stratified cross-validation with 10
data partitions is repeated in another step of class modeling,
parameter tuning and classification of test partition samples.
Despite the high specificity observed in Table 2, a fairly low sen-
sitivity can also be noted. In Section 1.1 the labels that serve as
ground truth for the evaluation of performance are explicitly
defined based on the pre-ignition addition of an IL to a sample. It
7
is possible that for some proportion of the samples, the IL is com-
pletely consumed during the course of the fire. This could lead to
cases where an sample labeled as containing ILR contains no resid-
ual ILR. Previous studies have chosen to define ground truth in
terms of labels assigned by an informed analyst (an analyst aware
of the presence or absence of a particular IL at the time of sample
ignition) [1618], however in this manuscript we opt to accept the
risk of a biased (high) FN rate in order to maintain an absolutely
objective labeling scheme, and therefore conservative performance
reporting.
3.2. Classification of ignitable liquid residues
Only samples that were predicted to contain ILR are passed on
to the classification step, however samples that were incorrectly
classified as not containing ILR may yet again be used in the train-
ing or parameter tuning partitions. Likewise, the intact IL samples
are now allowed in the training and parameter tuning partition in
order to encourage features distinctive to the differences between
ILR classes to be modeled. The underlying assumption is that a dif-
ferent set of features may be required to distinguish between ILR
classes than was required to detect them. A different Monte Carlo
simulation is used for the results reported for the classification
results reported in Table 3 in order to allow a greater number of
iterations and therefore a larger number of samples to pass from
the detection step to the classification step. A similar FP rate to
the one reported in Table 2 is seen with 80 FP seen out of the
7420 total samples (1.1%) for which classifications were
performed.
The overall successful classification of an ILR, assuming a pre-
liminary successful detection step, is 74.9% (5560 of the 7420 clas-
sifications performed). The proportion of samples classified in
Table 3 will not necessarily reflect the relative class frequencies
in the experimental design because there may be a systematic bias
in the preliminary detection step (i.e. some ILs are more readily
detected and therefore more frequently passed on to the classifica-
tion step). The results reported for the testing of classification
accuracy do not contain any intact liquid samples; only fire debris
is classified into the three classes based on the features that are
detectable after highly variable fire debris samples have been
burnt as per the protocol described in Section 2.1.
It is obvious that the GAS class is distinct within the projected
space defined by LDA as quite good classification is achieved for
samples containing gasoline ILR, whereas the LO and WS classes
are more frequently misclassified as one another, suggesting a pos-
sible chemical similarity in these classes.
8 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113

Fig. 5. Example chromatograms of the three IL classes examined in this research. (a) Typical GC
GC-MS chromatogram of a neat white spirit (WS) sample (b) Typical
GC
GC-MS chromatogram of a neat gasoline (GAS) sample (c) Typical chromatogram of a neat lamp oil (LO) sample.

Table 2 An additional observation made when examining specific cases

Performance observed for the detection of ignitable liquid residue, using the
likelihood ratio formulation with a threshold of 1:0.
of erroneously classified samples was that occasionally, new sam-
ples to be classified would appear very far away from all samples
used in the modeling set to define the distributions (as depicted
in Fig. 4). This may be an indication that the number of samples
used in this study is simply insufficient to achieve better classifica-
tion results given the complexity of the fire debris matrix used. As
with many statistical models relying on the approximation of class
characteristics in the form of a probability density function, perfor-
mance is expected to improve with increased sampling as the
 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
Table 3
Performance observed for the classification of ignitable liquid residues, using the maximum likelihood of class distributions modeled form training samples.
Fig. 6. Tippett plot of likelihood ratios observed for ignitable liquid detection.
diversity of the variation possible in samples is better captured and
the sample characteristics become more accurately modeled.
Another interesting observation is that the row sums for the
classification data (see Table 3) show a greater tendency for suc-
cessful detection of ILR in samples containing WS compared to
samples containing GAS. This results in a smaller proportion of
GAS samples being passed to the classification stage. This may be
related to the higher volatility of GAS compared to the other ILs,
where perhaps only GAS containing samples with a very strong
ILR compound signal are subject to classification. This observation
can only be made with the prior knowledge that the original data
set is balanced in terms of class representation.
3.3. Interpretation of evidential value
The use of the likelihood ratio framework for the assessment of
evidence allows a direct representation of the projection of new
points into a quantitative assessment of the evidential value. Due
to the dimension reduction performed by LDA, the models produce
a wide range of LRs for a certain type of comparison of samples. For
this reason it is critical to examine the overall behavior of the LR
system. A Tippet plot is generated from the aggregate results
reported in Table 2 and shown in Fig. 6.
A Tippett plot contains information about the proportions of LRs
observed in the system that fall above and below the point of neu-
tral evidence. The x-axis is plotted on a log scale, so 0 is the point at
which evidence is deemed uninformative in supporting either
hypothesis. The explanation for the very low observed FP rate is
immediately clear by the portion of the known negative ILR sam-
9
ples on the right side of the green (dashed) line. While this is a
desirable characteristic for a forensic context, the Tippett plot also
shows that the proportion of false negatives is very high. Subse-
quent classification performance will contain some additional
erroneous classifications of samples that do not contain ILR but
must be nonetheless classified into an ILR type.
We also note that the log10 LR values often suggest extremely
large evidential values. Given the size of the dataset and the diffi-
culty of this problem this may be a gross over estimation of the
quantitative evidential value. Tail regions of the modeled distribu-
tions produce values not adequately substantiated by the available
data and their ratios become unstable. More research, including
much more data collection and a full validation of the method
(including calibration of the LR values) is required before the
method can be implemented for casework.
As with the analysis of the ILR detection results, it is of interest
to monitor the behavior of the classification results in terms of the
LRs for competing class membership hypotheses. Performing a
one-versus-one examination allows the generation of a series of
Tippett plots similar to those shown in Fig. 6. In Fig. 7(a), (b), and
(c) the LRs observed for each particular ILR class are shown for
the competing hypotheses in Table 3. In each case two curves sum-
marize correct LRs, that is the LRs for models where the appropriate
class is in the numerator hypothesis (Hp ). A third line is presented
in each case showing the LRs observed when an ILR positive sam-
ple of a known class is compared against two inappropriate distri-
butions, related to the hypotheses of class membership in the two
remaining classes. These additional lines (plotted in black) should
yield LRs fairly balanced on either side of the neutral-evidence line.
10 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113

Fig. 7. Tippett plots for three ignitable liquid classes of interest demonstrating likelihood ratios observed from the evaluation of different competing hypotheses. (a) WS
samples evaluated against different hypotheses for classification.(b) GAS samples evaluated against different hypotheses for classification.(c) LO samples evaluated against
different hypotheses for classification.

These plots provide additional information pertaining to the mag-
nitudes of LRs observed and the cumulative proportion of LRs fall-
ing in that range for correct and erroneous evidential assertions.
The performance reported in Table 3 is well explained by the
LRs observed. Both sets of hypotheses for the GAS result in a high
proportion of LRs on the right side of the neutral evidence line in
Fig. 7(b). Samples from the WS class show similar LRs for both
denominator hypotheses and a proportion of LRs on either side
of the neutral evidence line as would be expected from the
reported classification performance. Interesting behavior is
observed for the LO samples, which are well differentiated from
GAS samples, obvious from the red line in Fig. 7(b). But this is
not the case when discriminating LO samples from WS samples,
as seen by the blue line in Fig. 7(c).
4. Discussion
4.1. Compounds of interest detected
Over the course of many iterations with intermediate random-
ization of sample order, it is possible to keep track of which
 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113 11

features in the RFR are most frequently selected. As mentioned in Table 4

Section 2.4, the feature selection retains full resolution in terms Compounds identified by statistical feature selection and EICs. An x indicates the
discovery of this compound in at least once sample ignited using that ignitible liquid.
of the nominal mass values and retains some spatial resolution
(within the tile region).
This allows us to depict the original chromatographic data in
terms of a Extracted Ion Chromatogram (EIC) wherein each tile
region is plotted using only the ions selected as useful for that
region. This figure is useful in exploring the identity of compounds
that are selected as useful in detection and classification of fire
debris samples based on their ILR content. Figure 9(b) shows the
result of such a plot. This visual tool is essential in forensic case-
work in order for the forensic chemist to understand the output
of the numerical methods and likewise for the statistical under-
standing of the chemical basis of evidential assertions.
One of the peaks observed in Fig. 8(b) is not even visible in the
corresponding location of Fig. 8(a). This compound was present in
very low abundance in this sample. Its apparent absence in the TIC
is due to the color scaling required to visualize the range of inten-
sities in either plot. The mass spectra of the peak can however be
isolated and are shown in the lower left hand corner of Figure (a).
Based on the mass spectrum this compound is tentatively
identified as a C3 substituted benzene (nominal molecular weight
120, molecular formula C9 H12 ). Compounds of this class are
expected to be present in most ILs.
The EICs generated for several samples were used to retrieve
the LIS which could then be passed as query spectra to the NIST/
EPA/NIH Mass Spectral Database (NIST 11). For each query spectra
the top 3 candidate compounds returned as library hits were man-
ually examined. To ensure confident identification all identified
compounds were searched for in the main library, using the NIST
MS Search program (NIST MS Search 2.0), the returned spectra determined useful in the classification of samples into different
were manually compared to the LIS spectra suspected to contain IL categories.
the candidate compound. Some example compound matches and
corresponding reference spectra from the NIST library can be seen 4.2. Future work
in the Supplementary Material.
Systematic examination of compounds regularly discovered in Preliminary results have been presented showing that a two-
samples known to have contained particular ignitable liquids in dimensional chromatographic separation allows the isolation of
the experimental design allows the identification of compounds valuable features for the classification and interpretation of fire
indicative of the presence of particular ignitable liquid classes. debris samples without having to directly address complications
Table 4 summarizes several compounds that were repeatedly related to peak integration. Current results are based on a coarse

Fig. 8. Comparison of the total ion chromatogram of a selected region in sample # 93 versus the extracted ion chromatogram showing only masses selected by the CLLR
feature selection method. (a) Total ion chromatogram (TIC) of sample # 93, zoomed in on the region of interest. (b) Extracted ion chromatogram (EIC) of sample #93, zoomed
in on the region of interest.
12 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
division of the chromatographic space into 12 tiles. An immediate
consideration for future work is the exploration of different strate-
gies for dividing the chromatographic space and computing LIS fea-
tures. Intuitively, the regions should be as small as possible to
include the least amount of interfering noise and the greatest pro-
portion of interesting analytes for the forensic problem. Conversely,
shrinking the tiles approaches a pixel by pixel comparison which
introduces problems with inter-sample alignment. This is the
objective of ongoing research.
In Section 4.1 an example is presented demonstrating the
potential assistance in identifying chemical characteristics of
forensic interest for the detection and classification of ILR in fire
debris. The result is achieved by visualizing the original analytical
data while applying different ion filters over the spatial regions
corresponding to the original division of the chromatogram. The
results presented here demonstrate the potential utility of this
strategy in assisting a forensic chemist with the examination of a
case sample. Currently underway is an extensive survey aimed at
providing an extensive analysis of chemical markers of potential
interest in ILR detection. This will be accomplished by using the
procedure depicted in Fig. 8(b) to search for additional trace com-
pounds of interest in other samples and determine if new indica-
tors of ILR can be discovered. Of additional interest would be the
future exploration of the evidential value of new markers in a tar-
geted approach for ILR detection and classification.
5. Conclusions
A novel method for performing detection and subsequent clas-
sification of ILR has been presented using GC
GCMS analyses of
fire debris samples. Preliminary findings demonstrate reasonable
detection rates (84%) with low FP occurrence (<1%) on a set of real-
istic fire debris samples. Current classification rates present a real-
istic prospective on the applicability of this method to automated
prescreening of fire debris samples in a forensic context.
A two-step evidence-based feature selection method is pre-
sented for GC
GCMS chromatograms, which demonstrates the
potential of this analytical technology for use in fire debris foren-
sics. The reduction from a very large dimensionality to a robust
and information rich reduced feature space is demonstrated. This
method takes advantage of the vast separation space while avoid-
ing the shortcomings typically encountered in GC
GC. The fea-
ture selection protocol provides intuitively interpretable selection
and discovery of pertinent chemical information, isolating com-
pounds in both the time domain and in the mass domain. This
allows experts to pinpoint the source of the characteristics in the
original chemical analysis that are the main basis for the evidential
assertions reported. By projection of the already reduced feature
set into a suitable space for classification via LDA, the desired size
of the reduced feature space itself is flexible and should theoreti-
cally yield better performance with the availability of larger foren-
sic reference collections.
The method also allows the calculation of, in a fully objective
manner, the evidential values (LRs) for comparing two hypotheses
of interest. Although promising results were obtained, further
research and extensive additional data collection is necessary for
casework implementation.
Acknowledgement
This research was funded by NWO Grant No. 727.011.006,
under the research objectives of the COMFOR project. The authors
acknowledge and appreciate the contribution of Brenda van Daelen
in providing laboratory assistance during sample preparation.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.forc.2016.10.003.
References
[1] CBS Netherlands. Brandweerstatistiek. Technical report, Centraal Bureau voor
de Statistiek, 01 2013, Language of Publication, Dutch.
[2] Jeanet Hendrikse, Michiel Grutters, Frank Schfer, Identifying Ignitable Liquids
in Fire Debris, A Guideline for Forensic Experts, Academic Press, October 2015.
[3] Maximilian K. Jennerwein, Markus Eschner, Thomas Grger, Thomas Wilharm,
Ralf Zimmermann, Complete group-type quantification of petroleum middle
distillates based on comprehensive two-dimensional gas chromatography
time-of-flight mass spectrometry (GC
GC-TOFMS) and visual basic scripting,
Energy Fuels 28 (9) (2014) 56705681.
[4] Eric Stauffer, Aromatic content in medium range distillate products, J. Forensic
Sci. 49 (5) (2004).
[5] Eric Stauffer, Doug Byron, Alternative fuels in fire debris analysis: Biodiesel
basics, J. Forensic Sci. 52 (2) (2007) 371379.
[6] Kaitlin R. Prather, Victoria L. McGuffin, Ruth Waddell Smith, Effect of
evaporation and matrix interferences on the association of simulated
ignitable liquid residues to the corresponding liquid standard, Forensic Sci.
Int. 222 (13) (2012) 242251.
[7] Peter Vergeer, Annabel Bolck, Leo J.C. Peschier, Charles E.H. Berger, Jeanet N.
Hendrikse, Likelihood ratio methods for forensic comparison of evaporated
gasoline residues, Sci. Justice 54 (6) (2014) 401411.
[8] Eric Stauffer, Concept of pyrolysis for fire debris analysts, Sci. Justice 43 (2003)
2940.
[9] M.S. Fernandes, C.M. Lau, W.C. Wong, The effect of volatile residues in burnt
household items on the detection of fire accelerants, Sci. Justice 42 (2001) 715.
[10] Patricia A. Contreras, Stephen S. Houck, William M. Davis, Jorn C.-C. Yu,
Pyrolysis products of linear alkylbenzenes implications in fire debris
analysis, J. Forensic Sci. 58 (1) (2013) 210216.
[11] Dee A. Turner, John V. Goodpaster, Comparing the effects of weathering and
microbial degradation on gasoline using principal components analysis, J.
Forensic Sci. 57 (1) (2012) 6469.
[12] D.A. Turner, J. Pichtel, Y. Rodenas, J. McKillip, J.V. Goodpaster, Microbial
degradation of gasoline in soil: effect of season of sampling, Forensic Sci. Int.
251 (2015) 6976.
[13] Jeanet Hendrikse, ENFSI collaborative testing programme for ignitable liquid
analysis: a review. Forensic Science International, 167(23):213219, 2007.
Selected Articles of the 4th European Academy of Forensic Science Conference
(EAFS2006) June 1316, 2006 Helsinki, Finland.
[14] Michael E. Sigman, Mary R. Williams, Rebecca J. Ivy, Individualization of
gasoline samples by covariance mapping and gas chromatography/mass
spectrometry, Anal. Chem. 79 (9) (2007) 34623468.
[15] Michael E. Sigman, Mary R. Williams, Joseph A. Castelbuono, Joseph G. Colca, C.
Douglas Clark, Ignitable liquid classification and identification using the
summed-ion mass spectrum, Instrum. Sci. Technol. 36 (4) (2008) 375393.
[16] Erin E. Waddell, E.T. Song, C.N. Rinke, Mary R. Williams, Michael E. Sigman,
Progress toward the determination of correct classification rates in fire debris
analysis, J. Forensic Sci. 58 (4) (2013) 887896.
[17] Erin E. Waddell, Mary R. Williams, Michael E. Sigman, Progress toward the
determination of correct classification rates in fire debris analysis ii: utilizing
soft independent modeling of class analogy (simca), J. Forensic Sci. 59 (2014)
927935.
[18] Martin Lopatka, Michael E. Sigman, Marjan J. Sjerps, Mary R. Williams, Gabriel
Viv-Truyols, Class-conditional feature modeling for ignitable liquid
classification with substantial substrate contribution in fire debris analysis,
Forensic Sci. Int. 252 (2015) 177186.
[19] Mary R. Williams, Michael E. Sigman, Jennifer Lewis, Kelly McHugh Pitan,
Combined target factor analysis and bayesian soft-classification of
interference-contaminated samples: Forensic fire debris analysis, Forensic
Sci. Int. 222 (13) (2012) 373386.
[20] A.S.T.M.E 1618, Standard Test Method for Ignitable Liquid Residues in Extracts
from Fire Debris Samples by Gas Chromatography-mass Spectrometry. ASTM,
2010.
[21] Eric Stauffer, Julia A. Dolan, Reta Newman, Fire Debris Analysis, first ed.,
Academic Press, 2008.
[22] Nikolai A. Sinkov, P.L. Mark Sandercock, James J. Harynuk, Chemometric
classification of casework arson samples based on gasoline content, Forensic
Sci. Int. 235 (2014) 2431.
[23] Andjoe Sampat, Martin Lopatka, Marjan Sjerps, Gabriel Viv-Truyols, Peter
Schoenmakers, and Arian van Asten. The forensic potential of comprehensive
two-dimensional gas chromatography. TrAC Trends in Analytical Chemistry,
2015.
[24] John J. Lentini, Review of: fire debris analysis, J. Forensic Sci. 53 (4) (2008)
1012.
[25] J.M. Raines, C.M. Taylor, A.K. Rosenhan, J.M. Rodriguez, An arson investigation
by using comprehensive two-dimensional gas chromatography-quadrupole
mass spectrometry, J. Forensic Res. 3 (2012).
[26] Kari L. Organtini, Anne L. Myers, Karl J. Jobst, Jack Cochran, Brian Ross, Brian
McCarry, Eric J. Reiner, Frank L. Dorman, Comprehensive characterization of the
 M. Lopatka et al. / Forensic Chemistry 3 (2017) 113
halogenated dibenzo-p-dioxin and dibenzofuran contents of residential fire
debris using comprehensive two-dimensional gas chromatography coupled to
time of flight mass spectrometry, J. Chromatogr. A 1369 (2014) 138146.
[27] Heather D. Bean, Jane E. Hill, Jean-Marie D. Dimandja, Improving the quality of
biomarker candidates in untargeted metabolomics via peak table-based
alignment of comprehensive two-dimensional gas chromatography-mass
spectrometry data, J. Chromatogr. A 1394 (2015) 111117.
[28] Hadi Parastar, Mehdi Jalali-Heravi, and Roma Tauler. Comprehensive two-
dimensional gas chromatography (GC
GC) retention time shift correction
and modeling using bilinear peak alignment, correlation optimized shifting
and multivariate curve resolution. Chemometrics and Intelligent Laboratory
Systems, 117:8091, 2012. Special Issue Section: Selected Papers from the 1st
African-European Conference on Chemometrics, Rabat, Morocco, September
2010 Special Issue Section: Preprocessing methods Special Issue Section:
Spectroscopic imaging.
[29] Martin Lopatka, Andrei Barcaru, Marjan J. Sjerps, Gabriel Viv-Truyols,
Leveraging probabilistic peak detection to estimate baseline drift in complex
chromatographic samples, J. Chromatogr. A 1431 (2016) 122130.
[30] M. Adahchour, J. Beens, U.A.Th. Brinkman, Recent developments in the
application of comprehensive two-dimensional gas chromatography, J.
Chromatogr. A 1186 (12) (2008) 67108, Trends and Developments in Gas
Chromatography.
[31] Jens Dallge, Jan Beens, Udo A.Th Brinkman, Comprehensive two-dimensional
gas chromatography: a powerful and versatile analytical tool, J. Chromatogr. A
1000 (1-2) (2003) 69108, A Century of Chromatography 19032003.
[32] Anders Nordgaard, Birgitta Rasmusson, The likelihood ratio as value of evidence
more than a question of numbers, Law Probab. Risk 11 (4) (2012) 303315.
[33] G. Zadora, A. Martyna, D. Ramos, C. Aitken, Statistical Analysis in Forensic
Science: Evidential Values of Multivariate Physicochemical Data, Wiley, 2013.
[34] Annabel Bolck, Haifang Ni, Martin Lopatka, Evaluating score- and feature-
based likelihood ratio models for multivariate continuous data: applied to
forensic mdma comparison, Law Probab. Risk 14 (3) (2015) 243266.
13
[35] C.G.G. Aitken, F. Taroni, Statistics and the Evaluation of Evidence for Forensic
Scientists. Statistics in Practice, Wiley, 2004.
[36] E. Beecher-Monas, Evaluating Scientific Evidence: An Interdisciplinary
Framework for Intellectual Due Process. Law in Context, Cambridge
University Press, 2006.
[37] Jorge Camargo, Pierre Esseiva, Fabio Gonzlez, Julien Wist, Luc Patiny,
Monitoring of illicit pill distribution networks using an image collection
exploration framework, Forensic Sci. Int. 223 (13) (2012) 298305.
[38] Kar-Weng Chan, Guan-Huat Tan, C.S. Richard, Wong. Investigation of trace
inorganic elements in street doses of heroin, Sci. Justice 53 (1) (2013) 7380.
[39] Amanda B. Hepler, Christopher P. Saunders, Linda J. Davis, JoAnn Buscaglia,
Score-based likelihood ratios for handwriting evidence, Forensic Sci. Int. 219
(13) (2012) 129140.
[40] T. Hicks, A. Biedermann, J.A. de Koeijer, F. Taroni, C. Champod, I.W. Evett, The
importance of distinguishing information from evidence/observations when
formulating propositions, Sci. Justice 55 (6) (2015) 520525.
[41] David A. Leeuwen, Niko Brmmer, Speaker classification i, in: Christian Mller
(Ed.), Speaker Classification I: Fundamentals, Features, and Methods, chapter
An Introduction to Application-Independent Evaluation of Speaker
Recognition Systems, Springer-Verlag, Berlin, Heidelberg, 2007, pp. 330353.
[42] Agnieszka Martyna, Grzegorz Zadora, Ivana Stanimirova, Daniel Ramos, Wine
authenticity verification as a forensic problem: an application of likelihood
ratio test to label verification, Food Chem. 150 (2014) 287295.
[43] Didier Meuwly, Forensic individualisation from biometric data, Sci. Justice 46
(4) (2006) 205213.
[44] T. Ns, A User-friendly Guide to Multivariate Calibration and Classification,
NIR Publications, 2002.
[45] H.S. Lee, B. Chen. Generalized likelihood ratio discriminant analysis, in:
Automatic Speech Recognition Understanding, 2009. ASRU 2009. IEEE
Workshop on, Nov 2009, pp. 158163.
[46] W.N. Venables, B.D. Ripley, Modern Applied Statistics with S, Springer, New
York, fourth edition, 2002.